An Adaptive Model Router can cut LLM inference costs by up to 90% in multi-agent systems without changing agent logic. Traditional designs assign the same powerful model — like GPT-5-Pro — to every agent (Planner, Researcher, Analyst, Reporter) regardless of task complexity. Static assignment fails when query types vary across departments or contexts, making runtime adaptability essential.

This architecture replaces a monolithic Global Planner with Just-In-Time (JIT) sub-task generation at each agent stage. When the Researcher acts, it invokes a Planner Agent scoped to data collection, using the actual upstream output to inform decisions. This eliminates context-blindness and reduces hallucination risk from massive upfront prompts.

A lightweight classification layer — powered by cheap models like Gemini-flash-lite or gpt-nano — evaluates each JIT task on the fly, routing it to the right-sized LLM. Classification is far simpler than execution, enabling accurate routing at minimal cost. The system provides granular visibility into inference spend while preserving agent autonomy and reasoning capability.